# Circuit-Level Detection

The cofiber threshold detection head expressed as synthesizable digital logic. All weights are analytically derived (zero gradient-based training) and stored as INT8 ROM.

## Person Detector (Proof of Concept)

Single-class (person) detector as a combinational circuit.

- **Parameters**: 4,614 (4.6 KB INT8 ROM)
- **Architecture**: 6 parallel dot products (1 cls + 4 reg + 1 ctr), each 768-dim
- **Weights**: closed-form least-squares on 20K COCO training images
- **Construction time**: 122 seconds, zero training steps

### Gate Count (Yosys synthesis)

Synthesized with Yosys (generic target, no FPGA mapping):

| Component | Dims | Gates (16-dim measured) | Gates (768-dim extrapolated) |
|-----------|------|------------------------|------------------------------|
| 1 MAC unit | 768→1 | 1,358 | ~65K |
| Person detector (6 MACs) | 768→{1,4,1} | — | ~391K |
| Full 80-class detector (85 MACs) | 768→{80,4,1} | — | ~7.1M |

### Files

- `person_detector.sv` — SystemVerilog module, Icarus Verilog compatible
- `tb_person.sv` — testbench
- `cofiber_detector.sv` — full 80-class detector (SystemVerilog, parameterized)
- `person_small_synth.v` — 16-dim reduced version for synthesis analysis
- `rom/person_cls_w.hex` — classification weights (768 INT8)
- `rom/person_reg_w.hex` — regression weights (4×768 INT8)
- `rom/person_ctr_w.hex` — centerness weights (768 INT8)
- `person_analytical.pth` — PyTorch checkpoint of the analytical solution

### Simulation

```bash
iverilog -g2012 -o tb_person.vvp person_detector.sv tb_person.sv
vvp tb_person.vvp
```

### Synthesis

```bash
yosys -p "read_verilog person_small_synth.v; synth -top person_detector_small; stat"
```

## Cofiber Decomposition

The spatial decomposition (avg_pool → subtract → repeat) is a fixed filter bank with zero learned parameters. In hardware, it is:
- Average pooling: 2×2 adder tree + arithmetic right shift
- Cofiber extraction: subtractor (input − upsampled pool)
- Fully combinational, no state

The decomposition module (`cofiber_detector.sv`) implements all three scales for a 40×40 input grid.